Level 2 · How deep do visitors browse?

What a histogram shows

A histogram splits a numeric column into equal ranges and draws one bar per range, as tall as the number of values in it. For pages viewed, each range is a single page count, so the first bar is “sessions that viewed 1 page”, the next “2 pages”, and so on. Reading one comes down to three questions: where is the peak, how wide is the spread, and is one side longer than the other?

web_sessions["pages_viewed"].value_counts().sort_index().head()

1    17405
2     4932
3     3635
4     3003
5     2368

Skew and the mean

When the tail stretches to the right, the distribution is right-skewed. The long tail pulls the mean toward it, while the median stays with the bulk of the data:

pages = web_sessions["pages_viewed"]
pages.mean(), pages.median(), pages.skew()   # 3.66, 2.0, 2.35

In a right-skewed column the mean is above the median; in a left-skewed one (rare in business data, an example is days until an expiry date) it is below. The bigger the gap, the less the mean describes a typical value.

Spikes mean a second story

A single bar far taller than its neighbours usually means a different process is mixed in. Here it’s bounces: visitors who land on one page and leave. They are 43.5% of all sessions. Mixing bouncers and browsers in one average describes neither group. Split them and each gets a useful number:

(pages == 1).mean()          # 0.435 bounce rate
pages[pages > 1].median()    # 4.0 pages for visits that go past the first page

The normal curve is not the default

Many statistics lessons start with the bell-shaped normal distribution, where mean and median match and about 95% of values sit within two standard deviations of the mean. Counts, money and durations in business data rarely look like that. A quick test: if the mean minus two standard deviations is impossible (a negative page count, a negative price), the data is not normal, and rules built on that shape don’t apply.

Common mistakes

• Reporting the mean without looking at the shape. Always plot the histogram first.
• Ignoring a spike. A tall bar at 0, 1 or a round number is often a separate behaviour (bounces, default values, free orders) worth its own analysis.
• Calling the tail “errors”. Long tails are normal in business data. Remove a value only when you know it is wrong, not because it is big.